Research-hardened two-agent coding loop (Lead + adversarial Evaluator) for Claude Code, with a portable driver for Codex, Cursor, OpenCode, Pi, Hermes, Gemini and more — agents communicate via markdown mailboxes
Agent Claude Code
Sonnet primary implementation worker for the trio loop. Executes one well-specified increment handed down by the Opus Lead, including substantive application logic, tests, and integration work.
Agent Claude Code
Opus adversarial evaluator of the duo loop. Verifies the Lead's iteration against PLAN.md's acceptance criteria by actually exercising the code, using Sonnet explorers for scoped reconnaissance and Sonnet implementors only when delegated verification support is needed. Writes VERDICT.md with SHIP / ITERATE (optionally…
Agent Claude Code
Opus lead of the duo loop — plans, delegates the main implementation pass to Sonnet builders, then reviews and corrects their work. Maintains PLAN.md and REPORT.md and owns the final result.
Agent Claude Code
Sonnet scoped-repair worker for the trio loop. Invoked on VERDICT: ITERATE scope=local: — fixes exactly the listed failure scope with no re-planning, refactoring, or scope expansion.
Agent Claude Code
Sonnet read-only explorer for the trio loop. Answers codebase questions for the Opus Lead/Evaluator — how things work, where things live, what conventions exist. Never modifies anything.
Agent
Native Trio coordinator that runs the fixed Scout, Lead, Builder, review, and Evaluator protocol.